算術問題解答システムにおける学習

抄録

This paper presents a learning method for automating knowledge acquisition in an arithmetic problem solver. Our arithmetic problem solver with learning mechanism called LAPS can solve many different arithmetic problems given in natural language without reprogramming. LAPS consists of a natural language processor, equation extractor, equation solver, rule generator, rule modifier and knowledgebase. In such a problem solver, a huge amount of knowledge is required such as, common knowledge to understand given problems and domain specific knowledge to extract equations. When a given problem cannot be solved because of a shortage of such knowledge, LAPS can acquire the knowledge through interaction with an outside teacher. Knowledge obtained from a teacher is represented in the form of rules. Such obtained knowledge is usually rather specific and specific knowledge should be generalized and refined in order to use it for solving problems. Since a problem solver generally does not have enough knowledge to explain a given problem, explanation based generalization is not applicable. On the other hand, the similarity based learning requires a negative instance or some other constraint in order to avoid overgeneralization. In our problem solver, generalization is made by using the constraint that the given problem should be solved. Our learning method is based on empirical similarity but can avoid overgeneralization without giving negative instances. That is, the validity of the generalized rules are confirmed by solving problems, although the similarity-based learning cannot assure it. Furthermore, in this generalization process, LAPS can improve its performance at problem solving by synthesizing the applied rules.